- Generated on Fri Apr 20 2012 16:46:40 for libeblearn by
1.7.4
|
libeblearn
|
00001 /*************************************************************************** 00002 * Copyright (C) 2008 by Yann LeCun and Pierre Sermanet * 00003 * yann@cs.nyu.edu, pierre.sermanet@gmail.com * 00004 * 00005 * Redistribution and use in source and binary forms, with or without 00006 * modification, are permitted provided that the following conditions are met: 00007 * * Redistributions of source code must retain the above copyright 00008 * notice, this list of conditions and the following disclaimer. 00009 * * Redistributions in binary form must reproduce the above copyright 00010 * notice, this list of conditions and the following disclaimer in the 00011 * documentation and/or other materials provided with the distribution. 00012 * * Redistribution under a license not approved by the Open Source 00013 * Initiative (http://www.opensource.org) must display the 00014 * following acknowledgement in all advertising material: 00015 * This product includes software developed at the Courant 00016 * Institute of Mathematical Sciences (http://cims.nyu.edu). 00017 * * The names of the authors may not be used to endorse or promote products 00018 * derived from this software without specific prior written permission. 00019 * 00020 * THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED 00021 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 00022 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 00023 * DISCLAIMED. IN NO EVENT SHALL ThE AUTHORS BE LIABLE FOR ANY 00024 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 00025 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 00026 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 00027 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 00028 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 00029 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 00030 ***************************************************************************/ 00031 00032 namespace ebl { 00033 00035 // net_cscscf 00036 00037 // the empty constructor (must call init afterwards) 00038 template <typename T, class Tstate> 00039 net_cscscf<T,Tstate>::net_cscscf() 00040 : layers<T,Tstate>(true) { 00041 // owns modules, responsible for deleting it 00042 } 00043 00044 template <typename T, class Tstate> 00045 net_cscscf<T,Tstate>::net_cscscf(parameter<T,Tstate> &prm, intg ini, intg inj, 00046 intg ki0, intg kj0, idx<intg> &tbl0, 00047 intg si0, intg sj0, 00048 intg ki1, intg kj1, idx<intg> &tbl1, 00049 intg si1, intg sj1, 00050 intg ki2, intg kj2, idx<intg> &tbl2, 00051 intg outthick, bool norm, bool mirror, 00052 bool tanh, bool shrink, bool diag) 00053 : layers<T,Tstate>(true) { 00054 // owns modules, responsible for deleting it 00055 init(prm, ini, inj, ki0, kj0, tbl0, si0, sj0, ki1, kj1, tbl1, 00056 si1, sj1, ki2, kj2, tbl2, outthick, norm, mirror, tanh, shrink, diag); 00057 } 00058 00059 template <typename T, class Tstate> 00060 net_cscscf<T,Tstate>::~net_cscscf() {} 00061 00062 template <typename T, class Tstate> 00063 void net_cscscf<T,Tstate>::init(parameter<T,Tstate> &prm, intg ini, intg inj, 00064 intg ki0, intg kj0, idx<intg> &tbl0, 00065 intg si0, intg sj0, intg ki1, intg kj1, 00066 idx<intg> &tbl1, intg si1, intg sj1, 00067 intg ki2, intg kj2, idx<intg> &tbl2, 00068 intg outthick, bool norm, bool mirror, 00069 bool tanh, bool shrink, bool diag) { 00070 // here we compute the thickness of the feature maps based on the 00071 // convolution tables. 00072 idx<intg> tblmax = tbl0.select(1, 1); 00073 intg thick0 = 1 + idx_max(tblmax); 00074 tblmax = tbl1.select(1, 1); 00075 intg thick1 = 1 + idx_max(tblmax); 00076 tblmax = tbl2.select(1, 1); 00077 intg thick2 = 1 + idx_max(tblmax); 00078 // convolution sizes 00079 idxdim ker0(ki0, kj0); 00080 idxdim ker1(ki1, kj1); 00081 idxdim ker2(ki2, kj2); 00082 idxdim stride(1, 1); 00083 // subsampling sizes 00084 idxdim sker0(si0, sj0); 00085 idxdim sker1(si1, sj1); 00086 00087 // layers was initialized with true so it owns the modules we give it, 00088 // we can add modules with "new". 00089 // we add convolutions (c), subsamplings (s), and full layers (f) 00090 // to form a c-s-c-s-c-f network. and we add state_idx in between 00091 // which serve as temporary buffer to hold the output of a module 00092 // and feed the input of the following module. 00093 00094 // convolution 00095 this->add_module(new convolution_module<T,Tstate> 00096 (&prm, ker0, stride, tbl0, "c0")); 00097 // bias 00098 this->add_module(new addc_module<T,Tstate>(&prm, thick0, "c0bias")); 00099 // non linearity 00100 if (shrink) 00101 this->add_module(new smooth_shrink_module<T,Tstate>(&prm, thick0)); 00102 else if (tanh) 00103 this->add_module(new tanh_module<T,Tstate>()); 00104 else 00105 this->add_module(new stdsigmoid_module<T,Tstate>()); 00106 // feature coefficents 00107 if (diag) 00108 this->add_module(new diag_module<T,Tstate>(&prm, thick0)); 00109 // absolute rectification + contrast normalization 00110 if (norm) { 00111 this->add_module(new abs_module<T,Tstate>()); 00112 this->add_module(new contrast_norm_module<T,Tstate> 00113 (ker0, thick0, mirror, true, false, NULL, "w0")); 00114 } 00115 // subsampling 00116 this->add_module(new subsampling_layer<T,Tstate>(&prm, thick0,sker0,sker0, tanh)); 00117 // convolution 00118 this->add_module(new convolution_module<T,Tstate> 00119 (&prm, ker1, stride, tbl1)); 00120 // bias 00121 this->add_module(new addc_module<T,Tstate>(&prm, thick1)); 00122 // non linearity 00123 if (shrink) 00124 this->add_module(new smooth_shrink_module<T,Tstate>(&prm, thick1)); 00125 else if (tanh) 00126 this->add_module(new tanh_module<T,Tstate>()); 00127 else 00128 this->add_module(new stdsigmoid_module<T,Tstate>()); 00129 // feature coefficents 00130 if (diag) 00131 this->add_module(new diag_module<T,Tstate>(&prm, thick1)); 00132 // absolute rectification + contrast normalization 00133 if (norm) { 00134 this->add_module(new abs_module<T,Tstate>()); 00135 this->add_module(new contrast_norm_module<T,Tstate> 00136 (ker1, thick1, mirror, true, false, NULL, "w1")); 00137 } 00138 // subsampling 00139 this->add_module(new subsampling_layer<T,Tstate>(&prm, thick1,sker1,sker1, tanh)); 00140 // convolution + bias + sigmoid 00141 this->add_module(new convolution_layer<T,Tstate> 00142 (&prm, ker2, stride, tbl2, tanh)); 00143 // full 00144 this->add_module(new full_layer<T,Tstate>(&prm, thick2, outthick, tanh)); 00145 } 00146 00148 // net_cscf 00149 00150 // the empty constructor (must call init afterwards) 00151 template <typename T, class Tstate> 00152 net_cscf<T,Tstate>::net_cscf() 00153 : layers<T,Tstate>(true) { 00154 // owns modules, responsible for deleting it 00155 } 00156 00157 template <typename T, class Tstate> 00158 net_cscf<T,Tstate>::net_cscf(parameter<T,Tstate> &prm, intg ini, intg inj, 00159 intg ki0, intg kj0, idx<intg> &tbl0, 00160 intg si0, intg sj0, 00161 intg ki1, intg kj1, idx<intg> &tbl1, 00162 intg outthick, bool norm, bool mirror, 00163 bool tanh, bool shrink, bool diag, bool lut_features, 00164 idx<T> *lut) 00165 : layers<T,Tstate>(true) { 00166 // owns modules, responsible for deleting it 00167 init(prm, ini, inj, ki0, kj0, tbl0, si0, sj0, ki1, kj1, tbl1, 00168 outthick, norm, mirror, tanh, shrink, diag, lut_features, lut); 00169 } 00170 00171 template <typename T, class Tstate> 00172 net_cscf<T,Tstate>::~net_cscf() {} 00173 00174 template <typename T, class Tstate> 00175 void net_cscf<T,Tstate>::init(parameter<T,Tstate> &prm, intg ini, intg inj, 00176 intg ki0, intg kj0, idx<intg> &tbl0, 00177 intg si0, intg sj0, intg ki1, intg kj1, 00178 idx<intg> &tbl1, intg outthick, bool norm, 00179 bool mirror, bool tanh, bool shrink, bool diag, 00180 bool lut_features, idx<T> *lut) { 00181 // here we compute the thickness of the feature maps based on the 00182 // convolution tables. 00183 idx<intg> tblmax = tbl0.select(1, 1); 00184 intg thick0 = 1 + idx_max(tblmax); 00185 tblmax = tbl1.select(1, 1); 00186 intg thick1 = 1 + idx_max(tblmax); 00187 // convolution sizes 00188 idxdim ker0(ki0, kj0); 00189 idxdim ker1(ki1, kj1); 00190 idxdim stride(1, 1); 00191 // subsampling sizes 00192 idxdim sker0(si0, sj0); 00193 00194 // layers was initialized with true so it owns the modules we give it, 00195 // we can add modules with "new". 00196 // we add convolutions (c), subsamplings (s), and full layers (f) 00197 // to form a c-s-c-s-c-f network. and we add state_idx in between 00198 // which serve as temporary buffer to hold the output of a module 00199 // and feed the input of the following module. 00200 00201 // convolution 00202 this->add_module(new convolution_module<T,Tstate>(&prm, ker0, stride, 00203 tbl0, "c0")); 00204 // bias 00205 this->add_module(new addc_module<T,Tstate>(&prm, thick0, "c0bias")); 00206 // non linearity 00207 if (shrink) 00208 this->add_module(new smooth_shrink_module<T,Tstate>(&prm, thick0)); 00209 else if (tanh) 00210 this->add_module(new tanh_module<T,Tstate>()); 00211 else 00212 this->add_module(new stdsigmoid_module<T,Tstate>()); 00213 // feature coefficents 00214 if (diag) 00215 this->add_module(new diag_module<T,Tstate>(&prm, thick0)); 00216 // absolute rectification + contrast normalization 00217 if (norm) { 00218 this->add_module(new abs_module<T,Tstate>()); 00219 this->add_module(new contrast_norm_module<T,Tstate>(ker0, thick0, mirror, NULL, 00220 "w0")); 00221 } 00222 // subsampling 00223 this->add_module(new subsampling_layer<T,Tstate>(&prm,thick0,sker0,sker0,tanh,"s0")); 00224 // convolution 00225 this->add_module(new convolution_module<T,Tstate>(&prm,ker1,stride, 00226 tbl1,"c1")); 00227 // bias 00228 this->add_module(new addc_module<T,Tstate>(&prm, thick1, "c1bias")); 00229 // non linearity 00230 if (shrink) 00231 this->add_module(new smooth_shrink_module<T,Tstate>(&prm, thick1)); 00232 else if (tanh) 00233 this->add_module(new tanh_module<T,Tstate>()); 00234 else 00235 this->add_module(new stdsigmoid_module<T,Tstate>()); 00236 // feature coefficents 00237 if (diag) 00238 this->add_module(new diag_module<T,Tstate>(&prm, thick1)); 00239 // absolute rectification + contrast normalization 00240 if (norm) { 00241 this->add_module(new abs_module<T,Tstate>()); 00242 this->add_module(new contrast_norm_module<T,Tstate>(ker1, thick1, mirror, NULL, 00243 "w1")); 00244 } 00245 if (lut_features) 00246 this->add_module(new range_lut_module<T,Tstate>(lut)); 00247 // full 00248 this->add_module(new full_layer<T,Tstate>(&prm, thick1, outthick, tanh, "f2")); 00249 00250 // // convolution 00251 // if (norm) // absolute rectification + contrast normalization 00252 // this->add_module(new convabsnorm_layer<T,Tstate>(&prm, ker0,stride,tbl0,mirror,tanh)); 00253 // else // old fashioned way 00254 // this->add_module(new convolution_layer<T,Tstate>(&prm, ker0, stride, tbl0, tanh)); 00255 // // subsampling 00256 // this->add_module(new subsampling_layer<T,Tstate>(&prm,thick0,sker0, sker0, tanh));, 00257 // // convolution 00258 // if (norm) // absolute rectification + contrast normalization 00259 // this->add_module(new convabsnorm_layer<T,Tstate>(&prm, ker1,stride,tbl1,mirror,tanh)); 00260 // else // old fashioned way 00261 // this->add_module(new convolution_layer<T,Tstate>(&prm, ker1, stride, tbl1, tanh)); 00262 // // full 00263 // add_last_module(new full_layer<T,Tstate>(&prm, thick1, outthick, tanh)); 00264 } 00265 00267 // net_cscc 00268 00269 // the empty constructor (must call init afterwards) 00270 template <typename T, class Tstate> 00271 net_cscc<T,Tstate>::net_cscc() 00272 : layers<T,Tstate>(true) { 00273 // owns modules, responsible for deleting it 00274 } 00275 00276 template <typename T, class Tstate> 00277 net_cscc<T,Tstate>::net_cscc(parameter<T,Tstate> &prm, intg ini, intg inj, 00278 intg ki0, intg kj0, idx<intg> &tbl0, 00279 intg si0, intg sj0, 00280 intg ki1, intg kj1, idx<intg> &tbl1, 00281 idx<intg> &tbl2, 00282 intg outthick, bool norm, bool mirror, 00283 bool tanh, bool shrink, bool diag) 00284 : layers<T,Tstate>(true) { 00285 // owns modules, responsible for deleting it 00286 init(prm, ini, inj, ki0, kj0, tbl0, si0, sj0, ki1, kj1, tbl1, 00287 tbl2, outthick, norm, mirror, tanh, shrink, diag); 00288 } 00289 00290 template <typename T, class Tstate> 00291 net_cscc<T,Tstate>::~net_cscc() {} 00292 00293 template <typename T, class Tstate> 00294 void net_cscc<T,Tstate>::init(parameter<T,Tstate> &prm, intg ini, intg inj, 00295 intg ki0, intg kj0, idx<intg> &tbl0, 00296 intg si0, intg sj0, intg ki1, intg kj1, 00297 idx<intg> &tbl1, idx<intg> &tbl2, 00298 intg outthick, bool norm, 00299 bool mirror, bool tanh, bool shrink, bool diag) { 00300 // here we compute the thickness of the feature maps based on the 00301 // convolution tables. 00302 idx<intg> tblmax = tbl0.select(1, 1); 00303 intg thick0 = 1 + idx_max(tblmax); 00304 tblmax = tbl1.select(1, 1); 00305 intg thick1 = 1 + idx_max(tblmax); 00306 // WARNING: those two numbers must be changed 00307 // when image-height/image-width change 00308 // TODO: add assertion test here? 00309 intg ki2 = (((ini - ki0 + 1) / si0) - ki1 + 1); 00310 intg kj2 = (((inj - kj0 + 1) / sj0) - kj1 + 1); 00311 00312 // convolution sizes 00313 idxdim ker0(ki0, kj0); 00314 idxdim ker1(ki1, kj1); 00315 idxdim ker2(ki2, kj2); 00316 idxdim stride(1, 1); 00317 // subsampling sizes 00318 idxdim sker0(si0, sj0); 00319 00320 // convolution 00321 this->add_module(new convolution_module<T,Tstate>(&prm,ker0,stride, 00322 tbl0,"c0")); 00323 // bias 00324 this->add_module(new addc_module<T,Tstate>(&prm, thick0, "c0bias")); 00325 // non linearity 00326 if (shrink) 00327 this->add_module(new smooth_shrink_module<T,Tstate>(&prm, thick0)); 00328 else if (tanh) 00329 this->add_module(new tanh_module<T,Tstate>()); 00330 else 00331 this->add_module(new stdsigmoid_module<T,Tstate>()); 00332 // feature coefficents 00333 if (diag) 00334 this->add_module(new diag_module<T,Tstate>(&prm, thick0)); 00335 // absolute rectification + contrast normalization 00336 if (norm) { 00337 this->add_module(new abs_module<T,Tstate>()); 00338 this->add_module(new contrast_norm_module<T,Tstate>(ker0, thick0, mirror, 00339 NULL, "w0")); 00340 } 00341 // subsampling 00342 this->add_module(new subsampling_layer<T,Tstate>(&prm, thick0, sker0, sker0, tanh, "s0")); 00343 // convolution 00344 this->add_module(new convolution_module<T,Tstate>(&prm,ker1,stride, 00345 tbl1,"c1")); 00346 // bias 00347 this->add_module(new addc_module<T,Tstate>(&prm, thick1, "c1bias")); 00348 // non linearity 00349 if (shrink) 00350 this->add_module(new smooth_shrink_module<T,Tstate>(&prm, thick1)); 00351 else if (tanh) 00352 this->add_module(new tanh_module<T,Tstate>()); 00353 else 00354 this->add_module(new stdsigmoid_module<T,Tstate>()); 00355 // feature coefficents 00356 if (diag) 00357 this->add_module(new diag_module<T,Tstate>(&prm, thick1)); 00358 // absolute rectification + contrast normalization 00359 if (norm) { 00360 this->add_module(new abs_module<T,Tstate>()); 00361 this->add_module(new contrast_norm_module<T,Tstate>(ker1, thick1, mirror, 00362 NULL, "w1")); 00363 } 00364 // convolution + bias + sigmoid 00365 this->add_module(new convolution_layer<T,Tstate>(&prm, ker2, stride, tbl2, tanh,"c2")); 00366 } 00367 00369 // net_cscsc 00370 00371 // the empty constructor (must call init afterwards) 00372 template <typename T, class Tstate> 00373 net_cscsc<T,Tstate>::net_cscsc() 00374 : layers<T,Tstate>(true) { // owns modules, responsible for deleting it 00375 } 00376 00377 template <typename T, class Tstate> 00378 net_cscsc<T,Tstate>::net_cscsc(parameter<T,Tstate> &prm, intg ini, intg inj, 00379 intg ki0, intg kj0, idx<intg> &tbl0, 00380 intg si0, intg sj0, 00381 intg ki1, intg kj1, idx<intg> &tbl1, 00382 intg si1, intg sj1, 00383 intg ki2, intg kj2, idx<intg> &tbl2, 00384 bool norm, bool mirror, bool tanh, 00385 bool shrink, bool diag, bool norm_pos) 00386 : layers<T,Tstate>(true) { // owns modules, responsible for deleting it 00387 init(prm, ini, inj, ki0, kj0, tbl0, si0, sj0, ki1, kj1, tbl1, 00388 si1, sj1, ki2, kj2, tbl2, norm, mirror, tanh, shrink, diag, norm_pos); 00389 } 00390 00391 template <typename T, class Tstate> 00392 net_cscsc<T,Tstate>::~net_cscsc() {} 00393 00394 template <typename T, class Tstate> 00395 void net_cscsc<T,Tstate>::init(parameter<T,Tstate> &prm, intg ini, intg inj, 00396 intg ki0, intg kj0, idx<intg> &tbl0, 00397 intg si0, intg sj0, intg ki1, intg kj1, 00398 idx<intg> &tbl1, intg si1, intg sj1, 00399 intg ki2, intg kj2, idx<intg> &tbl2, 00400 bool norm, bool mirror, bool tanh, 00401 bool shrink, bool diag, bool norm_pos) { 00402 // here we compute the thickness of the feature maps based on the 00403 // convolution tables. 00404 idx<intg> tblmax = tbl0.select(1, 1); 00405 intg thick0 = 1 + idx_max(tblmax); 00406 tblmax = tbl1.select(1, 1); 00407 intg thick1 = 1 + idx_max(tblmax); 00408 00409 // convolution sizes 00410 idxdim ker0(ki0, kj0); 00411 idxdim ker1(ki1, kj1); 00412 idxdim ker2(ki2, kj2); 00413 idxdim stride(1, 1); 00414 // subsampling sizes 00415 idxdim sker0(si0, sj0); 00416 idxdim sker1(si1, sj1); 00417 00418 // layers was initialized with true so it owns the modules we give it, 00419 // we can add modules with "new". 00420 // we add convolutions (c), subsamplings (s), and full layers (f) 00421 // to form a c-s-c-s-c-f network. and we add state_idx in between 00422 // which serve as temporary buffer to hold the output of a module 00423 // and feed the input of the following module. 00424 00425 if (norm) 00426 cout << "Using contrast normalization " << (norm_pos?"after":"before") 00427 << " subsampling. " << endl; 00428 // convolution 00429 this->add_module(new convolution_module<T,Tstate> 00430 (&prm, ker0, stride, tbl0, "c0")); 00431 // bias 00432 this->add_module(new addc_module<T,Tstate>(&prm, thick0, "c0bias")); 00433 // non linearity 00434 if (shrink) 00435 this->add_module(new smooth_shrink_module<T,Tstate>(&prm, thick0)); 00436 else if (tanh) 00437 this->add_module(new tanh_module<T,Tstate>()); 00438 else 00439 this->add_module(new stdsigmoid_module<T,Tstate>()); 00440 // feature coefficents 00441 if (diag) 00442 this->add_module(new diag_module<T,Tstate>(&prm, thick0)); 00443 // absolute rectification 00444 if (norm) 00445 this->add_module(new abs_module<T,Tstate>()); 00446 // contrast normalization (positioned before sub) 00447 if (norm && !norm_pos) 00448 this->add_module(new contrast_norm_module<T,Tstate> 00449 (ker0, thick0, mirror, true, false, NULL, "w0")); 00450 // subsampling 00451 this->add_module(new subsampling_layer<T,Tstate>(&prm, thick0, sker0, sker0, tanh, "s0")); 00452 // contrast normalization (positioned after sub) 00453 if (norm && norm_pos) 00454 this->add_module(new contrast_norm_module<T,Tstate> 00455 (ker1, thick0, mirror, true, false, NULL, "w0")); 00456 // convolution 00457 this->add_module(new convolution_module<T,Tstate> 00458 (&prm, ker1, stride,tbl1, "c1")); 00459 // bias 00460 this->add_module(new addc_module<T,Tstate>(&prm, thick1, "c1bias")); 00461 // non linearity 00462 if (shrink) 00463 this->add_module(new smooth_shrink_module<T,Tstate>(&prm, thick1)); 00464 else if (tanh) 00465 this->add_module(new tanh_module<T,Tstate>()); 00466 else 00467 this->add_module(new stdsigmoid_module<T,Tstate>()); 00468 // feature coefficents 00469 if (diag) 00470 this->add_module(new diag_module<T,Tstate>(&prm, thick1)); 00471 // absolute rectification 00472 if (norm) 00473 this->add_module(new abs_module<T,Tstate>()); 00474 // contrast normalization (position before sub) 00475 if (norm && !norm_pos) 00476 this->add_module(new contrast_norm_module<T,Tstate> 00477 (ker1, thick1, mirror, true, false, NULL, "w1")); 00478 // subsampling 00479 this->add_module(new subsampling_layer<T,Tstate>(&prm, thick1, sker1, sker1, tanh, "s1")); 00480 // contrast normalization (position after sub) 00481 if (norm && norm_pos) 00482 this->add_module(new contrast_norm_module<T,Tstate> 00483 (ker2, thick1, mirror, true, false, NULL, "w1")); 00484 // convolution + bias + sigmoid 00485 this->add_module(new convolution_layer<T,Tstate> 00486 (&prm, ker2, stride, tbl2, tanh,"c2")); 00487 } 00488 00490 // lenet_cscsc 00491 00492 template <typename T, class Tstate> 00493 lenet_cscsc<T,Tstate>:: 00494 lenet_cscsc(parameter<T,Tstate> &prm, intg image_height, intg image_width, 00495 intg ki0, intg kj0, intg si0, intg sj0, intg ki1, 00496 intg kj1, intg si1, intg sj1, 00497 intg output_size, bool norm, bool color, bool mirror, bool tanh, 00498 bool shrink, bool diag, bool norm_pos, idx<intg> *table0_, idx<intg> *table1_, 00499 idx<intg> *table2_) 00500 : net_cscsc<T,Tstate>() { 00501 idx<intg> table0, table1, table2; 00502 if (!color) { // use smaller tables 00503 table0 = full_table(1, 6); 00504 table1 = idx<intg>(60, 2); 00505 memcpy(table1.idx_ptr(), connection_table_6_16, 00506 table1.nelements() * sizeof (intg)); 00507 } else { // for color (assuming 3-layer input), use bigger tables 00508 table0 = idx<intg>(14, 2); 00509 intg tbl0[14][2] = {{0, 0}, {0, 1}, {0, 2}, {0, 3}, {1, 4}, {2, 4}, 00510 {1, 5}, {2, 5}, {0, 6}, {1, 6}, {2, 6}, {0, 7}, 00511 {1, 7}, {2, 7}}; 00512 memcpy(table0.idx_ptr(), tbl0, table0.nelements() * sizeof (intg)); 00513 table1 = idx<intg>(96, 2); 00514 memcpy(table1.idx_ptr(), connection_table_8_24, 00515 table1.nelements() * sizeof (intg)); 00516 } 00517 // overide default tables if defined 00518 if (table0_) 00519 table0 = *table0_; 00520 if (table1_) 00521 table1 = *table1_; 00522 if (table2_) 00523 table2 = *table2_; 00524 else { 00525 // determine max of previous table 00526 idx<intg> outtable = table1.select(1, 1); 00527 table2 = full_table(idx_max(outtable) + 1, output_size); 00528 cout << "Using a full table for table 2: " << idx_max(outtable) + 1 00529 << " -> " << output_size << endl; 00530 } 00531 00532 // WARNING: those two numbers must be changed 00533 // when image-height/image-width change 00534 // TODO: add assertion test here? 00535 intg ki2 = (((image_height - ki0 + 1) / si0) - ki1 + 1) / si1; 00536 intg kj2 = (((image_width - kj0 + 1) / sj0) - kj1 + 1) / sj1; 00537 00538 this->init(prm, image_height, image_width, ki0, kj0, table0, si0, sj0, 00539 ki1, kj1, table1, si1, sj1, ki2, kj2, table2, norm, mirror, 00540 tanh, shrink, diag, norm_pos); 00541 } 00542 00544 // lenet 00545 00546 template <typename T, class Tstate> 00547 lenet<T,Tstate>::lenet(parameter<T,Tstate> &prm, intg image_height, 00548 intg image_width, intg ki0, intg kj0, intg si0, 00549 intg sj0, intg ki1, intg kj1, intg si1, intg sj1, 00550 intg hid, intg output_size, bool norm, bool color, 00551 bool mirror, bool tanh, bool shrink, bool diag, 00552 idx<intg> *table0_, idx<intg> *table1_, 00553 idx<intg> *table2_) 00554 : net_cscscf<T,Tstate>() { 00555 idx<intg> table0, table1, table2; 00556 if (!color) { // use smaller tables 00557 table0 = full_table(1, 6); 00558 table1 = idx<intg>(60, 2); 00559 memcpy(table1.idx_ptr(), connection_table_6_16, 00560 table1.nelements() * sizeof (intg)); 00561 } else { // for color (assuming 3-layer input), use bigger tables 00562 table0 = idx<intg>(14, 2); 00563 intg tbl0[14][2] = {{0, 0}, {0, 1}, {0, 2}, {0, 3}, {1, 4}, {2, 4}, 00564 {1, 5}, {2, 5}, {0, 6}, {1, 6}, {2, 6}, {0, 7}, 00565 {1, 7}, {2, 7}}; 00566 memcpy(table0.idx_ptr(), tbl0, table0.nelements() * sizeof (intg)); 00567 table1 = idx<intg>(96, 2); 00568 memcpy(table1.idx_ptr(), connection_table_8_24, 00569 table1.nelements() * sizeof (intg)); 00570 } 00571 // overide default tables if defined 00572 if (table0_) 00573 table0 = *table0_; 00574 if (table1_) 00575 table1 = *table1_; 00576 // allocate a full table for table2 by default 00577 idx<intg> tb1 = table1.select(1, 1); 00578 intg max1 = idx_max(tb1) + 1; 00579 table2 = full_table(max1, hid); 00580 // override table2 if defined 00581 if (table2_) 00582 table2 = *table2_; 00583 00584 // WARNING: those two numbers must be changed 00585 // when image-height/image-width change 00586 // TODO: add assertion test here? 00587 intg ki2 = (((image_height - ki0 + 1) / si0) - ki1 + 1) / si1; 00588 intg kj2 = (((image_width - kj0 + 1) / sj0) - kj1 + 1) / sj1; 00589 00590 this->init(prm, image_height, image_width, ki0, kj0, table0, si0, sj0, 00591 ki1, kj1, table1, si1, sj1, ki2, kj2, table2, output_size, 00592 norm, mirror, tanh, shrink, diag); 00593 } 00594 00596 // lenet 00597 00598 template <typename T, class Tstate> 00599 lenet_cscf<T,Tstate>::lenet_cscf(parameter<T,Tstate> &prm, 00600 intg image_height, intg image_width, 00601 intg ki0, intg kj0, intg si0, intg sj0, 00602 intg ki1, intg kj1, intg output_size, 00603 bool norm, bool color, bool mirror, 00604 bool tanh, bool shrink, bool diag, idx<intg> *table0_, 00605 idx<intg> *table1_) 00606 : net_cscf<T,Tstate>() { 00607 idx<intg> table0, table1; 00608 if (!color) { // use smaller tables 00609 table0 = full_table(1, 6); 00610 table1 = idx<intg>(60, 2); 00611 memcpy(table1.idx_ptr(), connection_table_6_16, 00612 table1.nelements() * sizeof (intg)); 00613 } else { // for color (assuming 3-layer input), use bigger tables 00614 table0 = idx<intg>(14, 2); 00615 intg tbl0[14][2] = {{0, 0}, {0, 1}, {0, 2}, {0, 3}, {1, 4}, {2, 4}, 00616 {1, 5}, {2, 5}, {0, 6}, {1, 6}, {2, 6}, {0, 7}, 00617 {1, 7}, {2, 7}}; 00618 memcpy(table0.idx_ptr(), tbl0, table0.nelements() * sizeof (intg)); 00619 table1 = idx<intg>(96, 2); 00620 memcpy(table1.idx_ptr(), connection_table_8_24, 00621 table1.nelements() * sizeof (intg)); 00622 } 00623 if (table0_) 00624 table0 = *table0_; 00625 if (table1_) 00626 table1 = *table1_; 00627 00628 this->init(prm, image_height, image_width, ki0, kj0, table0, si0, sj0, 00629 ki1, kj1, table1, output_size, norm, mirror, tanh, shrink, diag); 00630 } 00631 00633 // lenet5 00634 00635 template <typename T, class Tstate> 00636 lenet5<T,Tstate>::lenet5(parameter<T,Tstate> &prm, intg image_height, 00637 intg image_width, 00638 intg ki0, intg kj0, intg si0, intg sj0, 00639 intg ki1, intg kj1, intg si1, intg sj1, 00640 intg hid, intg output_size, bool norm, bool mirror, 00641 bool tanh, bool shrink, bool diag) { 00642 idx<intg> table0 = full_table(1, 6); 00643 // TODO: add idx constructor taking pointer to data with dimensions 00644 // and copies it. 00645 // or if possible operator= taking the array in brackets? 00646 idx<intg> table1(60, 2); 00647 memcpy(table1.idx_ptr(), connection_table_6_16, 00648 table1.nelements() * sizeof (intg)); 00649 idx<intg> table2 = full_table(16, hid); 00650 00651 // WARNING: those two numbers must be changed 00652 // when image-height/image-width change 00653 // TODO: add assertion test here? 00654 intg ki2 = (((image_height - ki0 + 1) / si0) - ki1 + 1) / si1; 00655 intg kj2 = (((image_width - kj0 + 1) / sj0) - kj1 + 1) / sj1; 00656 00657 this->init(prm, image_height, image_width, ki0, kj0, table0, si0, sj0, 00658 ki1, kj1, table1, si1, sj1, ki2, kj2, table2, output_size, 00659 norm, mirror, tanh, shrink, diag); 00660 } 00661 00663 // lenet7 00664 00665 template <typename T, class Tstate> 00666 lenet7<T,Tstate>::lenet7(parameter<T,Tstate> &prm, intg image_height, 00667 intg image_width, intg output_size, bool norm, 00668 bool mirror, bool tanh, bool shrink, bool diag) { 00669 intg ki0 = 5, kj0 = 5; 00670 intg si0 = 4, sj0 = 4; 00671 intg ki1 = 6, kj1 = 6; 00672 intg si1 = 3, sj1 = 3; 00673 intg hid = 100; 00674 00675 idx<intg> table0 = full_table(1, 8); 00676 idx<intg> table1(96, 2); 00677 memcpy(table1.idx_ptr(), connection_table_8_24, 00678 table1.nelements() * sizeof (intg)); 00679 idx<intg> table2 = full_table(24, hid); 00680 00681 // WARNING: those two numbers must be changed 00682 // when image-height/image-width change 00683 // TODO: add assertion test here? 00684 intg ki2 = (((image_height - ki0 + 1) / si0) - ki1 + 1) / si1; 00685 intg kj2 = (((image_width - kj0 + 1) / sj0) - kj1 + 1) / sj1; 00686 00687 this->init(prm, image_height, image_width, ki0, kj0, table0, si0, sj0, 00688 ki1, kj1, table1, si1, sj1, ki2, kj2, table2, output_size, 00689 norm, mirror, tanh, shrink, diag); 00690 } 00691 00693 // lenet7_binocular 00694 00695 template <typename T, class Tstate> 00696 lenet7_binocular<T,Tstate>:: 00697 lenet7_binocular(parameter<T,Tstate> &prm, intg image_height, 00698 intg image_width, intg output_size, bool norm, bool mirror, 00699 bool tanh, bool shrink, bool diag) { 00700 intg ki0 = 5, kj0 = 5; 00701 intg si0 = 4, sj0 = 4; 00702 intg ki1 = 6, kj1 = 6; 00703 intg si1 = 3, sj1 = 3; 00704 intg hid = 100; 00705 00706 idx<intg> table0(12, 2); 00707 intg tbl0[12][2] = 00708 {{0, 0}, {0, 1}, {1, 2}, 00709 {1, 3}, {0, 4}, {1, 4}, 00710 {0, 5}, {1, 5}, {0, 6}, 00711 {1, 6}, {0, 7}, {1, 7}}; 00712 memcpy(table0.idx_ptr(), tbl0, table0.nelements() * sizeof (intg)); 00713 idx<intg> table1(96, 2); 00714 memcpy(table1.idx_ptr(), connection_table_8_24, 00715 table1.nelements() * sizeof (intg)); 00716 idx<intg> table2 = full_table(24, hid); 00717 00718 // WARNING: those two numbers must be changed 00719 // when image-height/image-width change 00720 // TODO: add assertion test here? 00721 intg ki2 = (((image_height - ki0 + 1) / si0) - ki1 + 1) / si1; 00722 intg kj2 = (((image_width - kj0 + 1) / sj0) - kj1 + 1) / sj1; 00723 00724 this->init(prm, image_height, image_width, ki0, kj0, table0, si0, sj0, 00725 ki1, kj1, table1, si1, sj1, ki2, kj2, table2, output_size, 00726 norm, mirror, tanh, shrink, diag); 00727 } 00728 00730 // supervised euclidean machine 00731 00732 template <typename Tdata, class Tlabel, class Tstate> 00733 supervised_euclidean_machine<Tdata,Tlabel,Tstate>:: 00734 supervised_euclidean_machine(module_1_1<Tdata,Tstate> &m, 00735 idx<Tdata> &t,idxdim &dims) 00736 : fc_ebm2<Tdata,Tstate,bbstate_idx<Tlabel>,Tstate> 00737 (m, fout, (ebm_2<Tstate,bbstate_idx<Tlabel>,Tstate>&)fcost), 00738 fcost(t), fout(dims) { 00739 } 00740 00741 template <typename Tdata, class Tlabel, class Tstate> 00742 supervised_euclidean_machine<Tdata,Tlabel,Tstate>:: 00743 ~supervised_euclidean_machine() { 00744 } 00745 00747 00748 } // end namespace ebl
1.7.4